DFSNet: A 3D Point Cloud Segmentation Network toward Trees Detection in an Orchard Scene.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Orchard management requires precise navigation and spraying capabilities for robots.
• Existing semantic segmentation networks struggle with the complexity of orchard environments.

Purpose of the Study:

• To propose a novel deep learning network, DFSNet, for accurate 3D point cloud segmentation in orchards.
• To enhance the performance of orchard management robots in tasks like autonomic navigation and precision spraying.

Main Methods:

• Developed DFSNet featuring a local feature aggregation (LFA) layer and a dynamic fusion segmentation (Fus-Seg) architecture.
• LFA layer uses positional encoders and multi-stage hierarchy for local pattern aggregation.
• Fus-Seg module learns a multi-embedding space for point tag formatting and feature mining.

Main Results:

• DFSNet achieved 89.43% accuracy and 74.05% mIoU on orchard field datasets.
• Outperformed PointNet, PointNet++, D-PointNet++, DGCNN, and Point-NN in accuracy and mIoU.
• Demonstrated significant improvements, with accuracy gains up to 11.73% and mIoU gains up to 28.19%.

Conclusions:

• DFSNet effectively captures more information from orchard scene point clouds.
• Provides more accurate point cloud segmentation results crucial for orchard management.
• The proposed network offers a significant advancement for robotic applications in agriculture.